Fix fine-tune weight mixture

This commit is contained in:
Matthew Honnibal 2017-08-21 14:15:29 -05:00
parent c10f63bf10
commit 80acbc5f1f
1 changed files with 6 additions and 10 deletions

View File

@ -370,24 +370,20 @@ def fine_tune(embedding, combine=None):
vecs, bp_vecs = embedding.begin_update(docs, drop=drop) vecs, bp_vecs = embedding.begin_update(docs, drop=drop)
flat_tokvecs = embedding.ops.flatten(tokvecs) flat_tokvecs = embedding.ops.flatten(tokvecs)
flat_vecs = embedding.ops.flatten(vecs) flat_vecs = embedding.ops.flatten(vecs)
alpha = model.mix
minus = 1-model.mix
output = embedding.ops.unflatten( output = embedding.ops.unflatten(
(alpha * flat_tokvecs + minus * flat_vecs), lengths) (model.mix[0] * flat_tokvecs + model.mix[1] * flat_vecs), lengths)
def fine_tune_bwd(d_output, sgd=None): def fine_tune_bwd(d_output, sgd=None):
flat_grad = model.ops.flatten(d_output) flat_grad = model.ops.flatten(d_output)
model.d_mix += flat_tokvecs.dot(flat_grad.T).sum() model.d_mix[0] += flat_tokvecs.dot(flat_grad.T).sum()
model.d_mix += 1-flat_vecs.dot(flat_grad.T).sum() model.d_mix[1] += flat_vecs.dot(flat_grad.T).sum()
bp_vecs([d_o * minus for d_o in d_output], sgd=sgd) bp_vecs([d_o * model.mix[1] for d_o in d_output], sgd=sgd)
d_output = [d_o * alpha for d_o in d_output]
sgd(model._mem.weights, model._mem.gradient, key=model.id) sgd(model._mem.weights, model._mem.gradient, key=model.id)
model.mix = model.ops.xp.minimum(model.mix, 1.0) return [d_o * model.mix[0] for d_o in d_output]
return d_output
return output, fine_tune_bwd return output, fine_tune_bwd
model = wrap(fine_tune_fwd, embedding) model = wrap(fine_tune_fwd, embedding)
model.mix = model._mem.add((model.id, 'mix'), (1,)) model.mix = model._mem.add((model.id, 'mix'), (2,))
model.mix.fill(0.5) model.mix.fill(0.5)
model.d_mix = model._mem.add_gradient((model.id, 'd_mix'), (model.id, 'mix')) model.d_mix = model._mem.add_gradient((model.id, 'd_mix'), (model.id, 'mix'))
return model return model